Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/nfsclient/nfs_bio.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * Rick Macklem at The University of Guelph. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_kdtrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/bio.h> 43 #include <sys/buf.h> 44 #include <sys/kernel.h> 45 #include <sys/mbuf.h> 46 #include <sys/mount.h> 47 #include <sys/proc.h> 48 #include <sys/vmmeter.h> 49 #include <sys/vnode.h> 50 51 #include <vm/vm.h> 52 #include <vm/vm_extern.h> 53 #include <vm/vm_page.h> 54 #include <vm/vm_object.h> 55 #include <vm/vm_pager.h> 56 #include <vm/vnode_pager.h> 57 58 #include <nfs/nfsproto.h> 59 #include <nfsclient/nfs.h> 60 #include <nfsclient/nfsmount.h> 61 #include <nfsclient/nfsnode.h> 62 #include <nfsclient/nfs_kdtrace.h> 63 64 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 65 struct thread *td); 66 static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 67 struct ucred *cred, int ioflag); 68 69 extern int nfs_directio_enable; 70 extern int nfs_directio_allow_mmap; 71 72 /* 73 * Vnode op for VM getpages. 74 */ 75 int 76 nfs_getpages(struct vop_getpages_args *ap) 77 { 78 int i, error, nextoff, size, toff, count, npages; 79 struct uio uio; 80 struct iovec iov; 81 vm_offset_t kva; 82 struct buf *bp; 83 struct vnode *vp; 84 struct thread *td; 85 struct ucred *cred; 86 struct nfsmount *nmp; 87 vm_object_t object; 88 vm_page_t *pages; 89 struct nfsnode *np; 90 91 vp = ap->a_vp; 92 np = VTONFS(vp); 93 td = curthread; /* XXX */ 94 cred = curthread->td_ucred; /* XXX */ 95 nmp = VFSTONFS(vp->v_mount); 96 pages = ap->a_m; 97 count = ap->a_count; 98 99 if ((object = vp->v_object) == NULL) { 100 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 101 return (VM_PAGER_ERROR); 102 } 103 104 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 105 mtx_lock(&np->n_mtx); 106 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 107 mtx_unlock(&np->n_mtx); 108 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 109 return (VM_PAGER_ERROR); 110 } else 111 mtx_unlock(&np->n_mtx); 112 } 113 114 mtx_lock(&nmp->nm_mtx); 115 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 116 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 117 mtx_unlock(&nmp->nm_mtx); 118 /* We'll never get here for v4, because we always have fsinfo */ 119 (void)nfs_fsinfo(nmp, vp, cred, td); 120 } else 121 mtx_unlock(&nmp->nm_mtx); 122 123 npages = btoc(count); 124 125 /* 126 * If the requested page is partially valid, just return it and 127 * allow the pager to zero-out the blanks. Partially valid pages 128 * can only occur at the file EOF. 129 */ 130 VM_OBJECT_LOCK(object); 131 if (pages[ap->a_reqpage]->valid != 0) { 132 vm_page_lock_queues(); 133 for (i = 0; i < npages; ++i) { 134 if (i != ap->a_reqpage) 135 vm_page_free(pages[i]); 136 } 137 vm_page_unlock_queues(); 138 VM_OBJECT_UNLOCK(object); 139 return (0); 140 } 141 VM_OBJECT_UNLOCK(object); 142 143 /* 144 * We use only the kva address for the buffer, but this is extremely 145 * convienient and fast. 146 */ 147 bp = getpbuf(&nfs_pbuf_freecnt); 148 149 kva = (vm_offset_t) bp->b_data; 150 pmap_qenter(kva, pages, npages); 151 PCPU_INC(cnt.v_vnodein); 152 PCPU_ADD(cnt.v_vnodepgsin, npages); 153 154 iov.iov_base = (caddr_t) kva; 155 iov.iov_len = count; 156 uio.uio_iov = &iov; 157 uio.uio_iovcnt = 1; 158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 159 uio.uio_resid = count; 160 uio.uio_segflg = UIO_SYSSPACE; 161 uio.uio_rw = UIO_READ; 162 uio.uio_td = td; 163 164 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 165 pmap_qremove(kva, npages); 166 167 relpbuf(bp, &nfs_pbuf_freecnt); 168 169 if (error && (uio.uio_resid == count)) { 170 nfs_printf("nfs_getpages: error %d\n", error); 171 VM_OBJECT_LOCK(object); 172 vm_page_lock_queues(); 173 for (i = 0; i < npages; ++i) { 174 if (i != ap->a_reqpage) 175 vm_page_free(pages[i]); 176 } 177 vm_page_unlock_queues(); 178 VM_OBJECT_UNLOCK(object); 179 return (VM_PAGER_ERROR); 180 } 181 182 /* 183 * Calculate the number of bytes read and validate only that number 184 * of bytes. Note that due to pending writes, size may be 0. This 185 * does not mean that the remaining data is invalid! 186 */ 187 188 size = count - uio.uio_resid; 189 VM_OBJECT_LOCK(object); 190 vm_page_lock_queues(); 191 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 192 vm_page_t m; 193 nextoff = toff + PAGE_SIZE; 194 m = pages[i]; 195 196 if (nextoff <= size) { 197 /* 198 * Read operation filled an entire page 199 */ 200 m->valid = VM_PAGE_BITS_ALL; 201 KASSERT(m->dirty == 0, 202 ("nfs_getpages: page %p is dirty", m)); 203 } else if (size > toff) { 204 /* 205 * Read operation filled a partial page. 206 */ 207 m->valid = 0; 208 vm_page_set_valid(m, 0, size - toff); 209 KASSERT(m->dirty == 0, 210 ("nfs_getpages: page %p is dirty", m)); 211 } else { 212 /* 213 * Read operation was short. If no error occured 214 * we may have hit a zero-fill section. We simply 215 * leave valid set to 0. 216 */ 217 ; 218 } 219 if (i != ap->a_reqpage) { 220 /* 221 * Whether or not to leave the page activated is up in 222 * the air, but we should put the page on a page queue 223 * somewhere (it already is in the object). Result: 224 * It appears that emperical results show that 225 * deactivating pages is best. 226 */ 227 228 /* 229 * Just in case someone was asking for this page we 230 * now tell them that it is ok to use. 231 */ 232 if (!error) { 233 if (m->oflags & VPO_WANTED) 234 vm_page_activate(m); 235 else 236 vm_page_deactivate(m); 237 vm_page_wakeup(m); 238 } else { 239 vm_page_free(m); 240 } 241 } 242 } 243 vm_page_unlock_queues(); 244 VM_OBJECT_UNLOCK(object); 245 return (0); 246 } 247 248 /* 249 * Vnode op for VM putpages. 250 */ 251 int 252 nfs_putpages(struct vop_putpages_args *ap) 253 { 254 struct uio uio; 255 struct iovec iov; 256 vm_offset_t kva; 257 struct buf *bp; 258 int iomode, must_commit, i, error, npages, count; 259 off_t offset; 260 int *rtvals; 261 struct vnode *vp; 262 struct thread *td; 263 struct ucred *cred; 264 struct nfsmount *nmp; 265 struct nfsnode *np; 266 vm_page_t *pages; 267 268 vp = ap->a_vp; 269 np = VTONFS(vp); 270 td = curthread; /* XXX */ 271 /* Set the cred to n_writecred for the write rpcs. */ 272 if (np->n_writecred != NULL) 273 cred = crhold(np->n_writecred); 274 else 275 cred = crhold(curthread->td_ucred); /* XXX */ 276 nmp = VFSTONFS(vp->v_mount); 277 pages = ap->a_m; 278 count = ap->a_count; 279 rtvals = ap->a_rtvals; 280 npages = btoc(count); 281 offset = IDX_TO_OFF(pages[0]->pindex); 282 283 mtx_lock(&nmp->nm_mtx); 284 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 285 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 286 mtx_unlock(&nmp->nm_mtx); 287 (void)nfs_fsinfo(nmp, vp, cred, td); 288 } else 289 mtx_unlock(&nmp->nm_mtx); 290 291 mtx_lock(&np->n_mtx); 292 if (nfs_directio_enable && !nfs_directio_allow_mmap && 293 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 294 mtx_unlock(&np->n_mtx); 295 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 296 mtx_lock(&np->n_mtx); 297 } 298 299 for (i = 0; i < npages; i++) 300 rtvals[i] = VM_PAGER_ERROR; 301 302 /* 303 * When putting pages, do not extend file past EOF. 304 */ 305 if (offset + count > np->n_size) { 306 count = np->n_size - offset; 307 if (count < 0) 308 count = 0; 309 } 310 mtx_unlock(&np->n_mtx); 311 312 /* 313 * We use only the kva address for the buffer, but this is extremely 314 * convienient and fast. 315 */ 316 bp = getpbuf(&nfs_pbuf_freecnt); 317 318 kva = (vm_offset_t) bp->b_data; 319 pmap_qenter(kva, pages, npages); 320 PCPU_INC(cnt.v_vnodeout); 321 PCPU_ADD(cnt.v_vnodepgsout, count); 322 323 iov.iov_base = (caddr_t) kva; 324 iov.iov_len = count; 325 uio.uio_iov = &iov; 326 uio.uio_iovcnt = 1; 327 uio.uio_offset = offset; 328 uio.uio_resid = count; 329 uio.uio_segflg = UIO_SYSSPACE; 330 uio.uio_rw = UIO_WRITE; 331 uio.uio_td = td; 332 333 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 334 iomode = NFSV3WRITE_UNSTABLE; 335 else 336 iomode = NFSV3WRITE_FILESYNC; 337 338 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 339 crfree(cred); 340 341 pmap_qremove(kva, npages); 342 relpbuf(bp, &nfs_pbuf_freecnt); 343 344 if (!error) { 345 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 346 if (must_commit) { 347 nfs_clearcommit(vp->v_mount); 348 } 349 } 350 return rtvals[0]; 351 } 352 353 /* 354 * For nfs, cache consistency can only be maintained approximately. 355 * Although RFC1094 does not specify the criteria, the following is 356 * believed to be compatible with the reference port. 357 * For nfs: 358 * If the file's modify time on the server has changed since the 359 * last read rpc or you have written to the file, 360 * you may have lost data cache consistency with the 361 * server, so flush all of the file's data out of the cache. 362 * Then force a getattr rpc to ensure that you have up to date 363 * attributes. 364 * NB: This implies that cache data can be read when up to 365 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 366 * attributes this could be forced by setting n_attrstamp to 0 before 367 * the VOP_GETATTR() call. 368 */ 369 static inline int 370 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 371 { 372 int error = 0; 373 struct vattr vattr; 374 struct nfsnode *np = VTONFS(vp); 375 int old_lock; 376 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 377 378 /* 379 * Grab the exclusive lock before checking whether the cache is 380 * consistent. 381 * XXX - We can make this cheaper later (by acquiring cheaper locks). 382 * But for now, this suffices. 383 */ 384 old_lock = nfs_upgrade_vnlock(vp); 385 if (vp->v_iflag & VI_DOOMED) { 386 nfs_downgrade_vnlock(vp, old_lock); 387 return (EBADF); 388 } 389 390 mtx_lock(&np->n_mtx); 391 if (np->n_flag & NMODIFIED) { 392 mtx_unlock(&np->n_mtx); 393 if (vp->v_type != VREG) { 394 if (vp->v_type != VDIR) 395 panic("nfs: bioread, not dir"); 396 (nmp->nm_rpcops->nr_invaldir)(vp); 397 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 398 if (error) 399 goto out; 400 } 401 np->n_attrstamp = 0; 402 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 403 error = VOP_GETATTR(vp, &vattr, cred); 404 if (error) 405 goto out; 406 mtx_lock(&np->n_mtx); 407 np->n_mtime = vattr.va_mtime; 408 mtx_unlock(&np->n_mtx); 409 } else { 410 mtx_unlock(&np->n_mtx); 411 error = VOP_GETATTR(vp, &vattr, cred); 412 if (error) 413 return (error); 414 mtx_lock(&np->n_mtx); 415 if ((np->n_flag & NSIZECHANGED) 416 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 417 mtx_unlock(&np->n_mtx); 418 if (vp->v_type == VDIR) 419 (nmp->nm_rpcops->nr_invaldir)(vp); 420 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 421 if (error) 422 goto out; 423 mtx_lock(&np->n_mtx); 424 np->n_mtime = vattr.va_mtime; 425 np->n_flag &= ~NSIZECHANGED; 426 } 427 mtx_unlock(&np->n_mtx); 428 } 429 out: 430 nfs_downgrade_vnlock(vp, old_lock); 431 return error; 432 } 433 434 /* 435 * Vnode op for read using bio 436 */ 437 int 438 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 439 { 440 struct nfsnode *np = VTONFS(vp); 441 int biosize, i; 442 struct buf *bp, *rabp; 443 struct thread *td; 444 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 445 daddr_t lbn, rabn; 446 off_t end; 447 int bcount; 448 int seqcount; 449 int nra, error = 0, n = 0, on = 0; 450 451 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode")); 452 if (uio->uio_resid == 0) 453 return (0); 454 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 455 return (EINVAL); 456 td = uio->uio_td; 457 458 mtx_lock(&nmp->nm_mtx); 459 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 460 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 461 mtx_unlock(&nmp->nm_mtx); 462 (void)nfs_fsinfo(nmp, vp, cred, td); 463 } else 464 mtx_unlock(&nmp->nm_mtx); 465 466 end = uio->uio_offset + uio->uio_resid; 467 if (vp->v_type != VDIR && 468 (end > nmp->nm_maxfilesize || end < uio->uio_offset)) 469 return (EFBIG); 470 471 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 472 /* No caching/ no readaheads. Just read data into the user buffer */ 473 return nfs_readrpc(vp, uio, cred); 474 475 biosize = vp->v_bufobj.bo_bsize; 476 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 477 478 error = nfs_bioread_check_cons(vp, td, cred); 479 if (error) 480 return error; 481 482 do { 483 u_quad_t nsize; 484 485 mtx_lock(&np->n_mtx); 486 nsize = np->n_size; 487 mtx_unlock(&np->n_mtx); 488 489 switch (vp->v_type) { 490 case VREG: 491 nfsstats.biocache_reads++; 492 lbn = uio->uio_offset / biosize; 493 on = uio->uio_offset & (biosize - 1); 494 495 /* 496 * Start the read ahead(s), as required. 497 */ 498 if (nmp->nm_readahead > 0) { 499 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 500 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 501 rabn = lbn + 1 + nra; 502 if (incore(&vp->v_bufobj, rabn) == NULL) { 503 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 504 if (!rabp) { 505 error = nfs_sigintr(nmp, td); 506 return (error ? error : EINTR); 507 } 508 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 509 rabp->b_flags |= B_ASYNC; 510 rabp->b_iocmd = BIO_READ; 511 vfs_busy_pages(rabp, 0); 512 if (nfs_asyncio(nmp, rabp, cred, td)) { 513 rabp->b_flags |= B_INVAL; 514 rabp->b_ioflags |= BIO_ERROR; 515 vfs_unbusy_pages(rabp); 516 brelse(rabp); 517 break; 518 } 519 } else { 520 brelse(rabp); 521 } 522 } 523 } 524 } 525 526 /* Note that bcount is *not* DEV_BSIZE aligned. */ 527 bcount = biosize; 528 if ((off_t)lbn * biosize >= nsize) { 529 bcount = 0; 530 } else if ((off_t)(lbn + 1) * biosize > nsize) { 531 bcount = nsize - (off_t)lbn * biosize; 532 } 533 bp = nfs_getcacheblk(vp, lbn, bcount, td); 534 535 if (!bp) { 536 error = nfs_sigintr(nmp, td); 537 return (error ? error : EINTR); 538 } 539 540 /* 541 * If B_CACHE is not set, we must issue the read. If this 542 * fails, we return an error. 543 */ 544 545 if ((bp->b_flags & B_CACHE) == 0) { 546 bp->b_iocmd = BIO_READ; 547 vfs_busy_pages(bp, 0); 548 error = nfs_doio(vp, bp, cred, td); 549 if (error) { 550 brelse(bp); 551 return (error); 552 } 553 } 554 555 /* 556 * on is the offset into the current bp. Figure out how many 557 * bytes we can copy out of the bp. Note that bcount is 558 * NOT DEV_BSIZE aligned. 559 * 560 * Then figure out how many bytes we can copy into the uio. 561 */ 562 563 n = 0; 564 if (on < bcount) 565 n = min((unsigned)(bcount - on), uio->uio_resid); 566 break; 567 case VLNK: 568 nfsstats.biocache_readlinks++; 569 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 570 if (!bp) { 571 error = nfs_sigintr(nmp, td); 572 return (error ? error : EINTR); 573 } 574 if ((bp->b_flags & B_CACHE) == 0) { 575 bp->b_iocmd = BIO_READ; 576 vfs_busy_pages(bp, 0); 577 error = nfs_doio(vp, bp, cred, td); 578 if (error) { 579 bp->b_ioflags |= BIO_ERROR; 580 brelse(bp); 581 return (error); 582 } 583 } 584 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 585 on = 0; 586 break; 587 case VDIR: 588 nfsstats.biocache_readdirs++; 589 if (np->n_direofoffset 590 && uio->uio_offset >= np->n_direofoffset) { 591 return (0); 592 } 593 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 594 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 595 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 596 if (!bp) { 597 error = nfs_sigintr(nmp, td); 598 return (error ? error : EINTR); 599 } 600 if ((bp->b_flags & B_CACHE) == 0) { 601 bp->b_iocmd = BIO_READ; 602 vfs_busy_pages(bp, 0); 603 error = nfs_doio(vp, bp, cred, td); 604 if (error) { 605 brelse(bp); 606 } 607 while (error == NFSERR_BAD_COOKIE) { 608 (nmp->nm_rpcops->nr_invaldir)(vp); 609 error = nfs_vinvalbuf(vp, 0, td, 1); 610 /* 611 * Yuck! The directory has been modified on the 612 * server. The only way to get the block is by 613 * reading from the beginning to get all the 614 * offset cookies. 615 * 616 * Leave the last bp intact unless there is an error. 617 * Loop back up to the while if the error is another 618 * NFSERR_BAD_COOKIE (double yuch!). 619 */ 620 for (i = 0; i <= lbn && !error; i++) { 621 if (np->n_direofoffset 622 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 623 return (0); 624 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 625 if (!bp) { 626 error = nfs_sigintr(nmp, td); 627 return (error ? error : EINTR); 628 } 629 if ((bp->b_flags & B_CACHE) == 0) { 630 bp->b_iocmd = BIO_READ; 631 vfs_busy_pages(bp, 0); 632 error = nfs_doio(vp, bp, cred, td); 633 /* 634 * no error + B_INVAL == directory EOF, 635 * use the block. 636 */ 637 if (error == 0 && (bp->b_flags & B_INVAL)) 638 break; 639 } 640 /* 641 * An error will throw away the block and the 642 * for loop will break out. If no error and this 643 * is not the block we want, we throw away the 644 * block and go for the next one via the for loop. 645 */ 646 if (error || i < lbn) 647 brelse(bp); 648 } 649 } 650 /* 651 * The above while is repeated if we hit another cookie 652 * error. If we hit an error and it wasn't a cookie error, 653 * we give up. 654 */ 655 if (error) 656 return (error); 657 } 658 659 /* 660 * If not eof and read aheads are enabled, start one. 661 * (You need the current block first, so that you have the 662 * directory offset cookie of the next block.) 663 */ 664 if (nmp->nm_readahead > 0 && 665 (bp->b_flags & B_INVAL) == 0 && 666 (np->n_direofoffset == 0 || 667 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 668 incore(&vp->v_bufobj, lbn + 1) == NULL) { 669 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 670 if (rabp) { 671 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 672 rabp->b_flags |= B_ASYNC; 673 rabp->b_iocmd = BIO_READ; 674 vfs_busy_pages(rabp, 0); 675 if (nfs_asyncio(nmp, rabp, cred, td)) { 676 rabp->b_flags |= B_INVAL; 677 rabp->b_ioflags |= BIO_ERROR; 678 vfs_unbusy_pages(rabp); 679 brelse(rabp); 680 } 681 } else { 682 brelse(rabp); 683 } 684 } 685 } 686 /* 687 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 688 * chopped for the EOF condition, we cannot tell how large 689 * NFS directories are going to be until we hit EOF. So 690 * an NFS directory buffer is *not* chopped to its EOF. Now, 691 * it just so happens that b_resid will effectively chop it 692 * to EOF. *BUT* this information is lost if the buffer goes 693 * away and is reconstituted into a B_CACHE state ( due to 694 * being VMIO ) later. So we keep track of the directory eof 695 * in np->n_direofoffset and chop it off as an extra step 696 * right here. 697 */ 698 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 699 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 700 n = np->n_direofoffset - uio->uio_offset; 701 break; 702 default: 703 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 704 bp = NULL; 705 break; 706 }; 707 708 if (n > 0) { 709 error = uiomove(bp->b_data + on, (int)n, uio); 710 } 711 if (vp->v_type == VLNK) 712 n = 0; 713 if (bp != NULL) 714 brelse(bp); 715 } while (error == 0 && uio->uio_resid > 0 && n > 0); 716 return (error); 717 } 718 719 /* 720 * The NFS write path cannot handle iovecs with len > 1. So we need to 721 * break up iovecs accordingly (restricting them to wsize). 722 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 723 * For the ASYNC case, 2 copies are needed. The first a copy from the 724 * user buffer to a staging buffer and then a second copy from the staging 725 * buffer to mbufs. This can be optimized by copying from the user buffer 726 * directly into mbufs and passing the chain down, but that requires a 727 * fair amount of re-working of the relevant codepaths (and can be done 728 * later). 729 */ 730 static int 731 nfs_directio_write(vp, uiop, cred, ioflag) 732 struct vnode *vp; 733 struct uio *uiop; 734 struct ucred *cred; 735 int ioflag; 736 { 737 int error; 738 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 739 struct thread *td = uiop->uio_td; 740 int size; 741 int wsize; 742 743 mtx_lock(&nmp->nm_mtx); 744 wsize = nmp->nm_wsize; 745 mtx_unlock(&nmp->nm_mtx); 746 if (ioflag & IO_SYNC) { 747 int iomode, must_commit; 748 struct uio uio; 749 struct iovec iov; 750 do_sync: 751 while (uiop->uio_resid > 0) { 752 size = min(uiop->uio_resid, wsize); 753 size = min(uiop->uio_iov->iov_len, size); 754 iov.iov_base = uiop->uio_iov->iov_base; 755 iov.iov_len = size; 756 uio.uio_iov = &iov; 757 uio.uio_iovcnt = 1; 758 uio.uio_offset = uiop->uio_offset; 759 uio.uio_resid = size; 760 uio.uio_segflg = UIO_USERSPACE; 761 uio.uio_rw = UIO_WRITE; 762 uio.uio_td = td; 763 iomode = NFSV3WRITE_FILESYNC; 764 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 765 &iomode, &must_commit); 766 KASSERT((must_commit == 0), 767 ("nfs_directio_write: Did not commit write")); 768 if (error) 769 return (error); 770 uiop->uio_offset += size; 771 uiop->uio_resid -= size; 772 if (uiop->uio_iov->iov_len <= size) { 773 uiop->uio_iovcnt--; 774 uiop->uio_iov++; 775 } else { 776 uiop->uio_iov->iov_base = 777 (char *)uiop->uio_iov->iov_base + size; 778 uiop->uio_iov->iov_len -= size; 779 } 780 } 781 } else { 782 struct uio *t_uio; 783 struct iovec *t_iov; 784 struct buf *bp; 785 786 /* 787 * Break up the write into blocksize chunks and hand these 788 * over to nfsiod's for write back. 789 * Unfortunately, this incurs a copy of the data. Since 790 * the user could modify the buffer before the write is 791 * initiated. 792 * 793 * The obvious optimization here is that one of the 2 copies 794 * in the async write path can be eliminated by copying the 795 * data here directly into mbufs and passing the mbuf chain 796 * down. But that will require a fair amount of re-working 797 * of the code and can be done if there's enough interest 798 * in NFS directio access. 799 */ 800 while (uiop->uio_resid > 0) { 801 size = min(uiop->uio_resid, wsize); 802 size = min(uiop->uio_iov->iov_len, size); 803 bp = getpbuf(&nfs_pbuf_freecnt); 804 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 805 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 806 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 807 t_iov->iov_len = size; 808 t_uio->uio_iov = t_iov; 809 t_uio->uio_iovcnt = 1; 810 t_uio->uio_offset = uiop->uio_offset; 811 t_uio->uio_resid = size; 812 t_uio->uio_segflg = UIO_SYSSPACE; 813 t_uio->uio_rw = UIO_WRITE; 814 t_uio->uio_td = td; 815 KASSERT(uiop->uio_segflg == UIO_USERSPACE || 816 uiop->uio_segflg == UIO_SYSSPACE, 817 ("nfs_directio_write: Bad uio_segflg")); 818 if (uiop->uio_segflg == UIO_USERSPACE) { 819 error = copyin(uiop->uio_iov->iov_base, 820 t_iov->iov_base, size); 821 if (error != 0) 822 goto err_free; 823 } else 824 /* 825 * UIO_SYSSPACE may never happen, but handle 826 * it just in case it does. 827 */ 828 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 829 size); 830 bp->b_flags |= B_DIRECT; 831 bp->b_iocmd = BIO_WRITE; 832 if (cred != NOCRED) { 833 crhold(cred); 834 bp->b_wcred = cred; 835 } else 836 bp->b_wcred = NOCRED; 837 bp->b_caller1 = (void *)t_uio; 838 bp->b_vp = vp; 839 error = nfs_asyncio(nmp, bp, NOCRED, td); 840 err_free: 841 if (error) { 842 free(t_iov->iov_base, M_NFSDIRECTIO); 843 free(t_iov, M_NFSDIRECTIO); 844 free(t_uio, M_NFSDIRECTIO); 845 bp->b_vp = NULL; 846 relpbuf(bp, &nfs_pbuf_freecnt); 847 if (error == EINTR) 848 return (error); 849 goto do_sync; 850 } 851 uiop->uio_offset += size; 852 uiop->uio_resid -= size; 853 if (uiop->uio_iov->iov_len <= size) { 854 uiop->uio_iovcnt--; 855 uiop->uio_iov++; 856 } else { 857 uiop->uio_iov->iov_base = 858 (char *)uiop->uio_iov->iov_base + size; 859 uiop->uio_iov->iov_len -= size; 860 } 861 } 862 } 863 return (0); 864 } 865 866 /* 867 * Vnode op for write using bio 868 */ 869 int 870 nfs_write(struct vop_write_args *ap) 871 { 872 int biosize; 873 struct uio *uio = ap->a_uio; 874 struct thread *td = uio->uio_td; 875 struct vnode *vp = ap->a_vp; 876 struct nfsnode *np = VTONFS(vp); 877 struct ucred *cred = ap->a_cred; 878 int ioflag = ap->a_ioflag; 879 struct buf *bp; 880 struct vattr vattr; 881 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 882 daddr_t lbn; 883 off_t end; 884 int bcount; 885 int n, on, error = 0; 886 887 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode")); 888 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 889 ("nfs_write proc")); 890 if (vp->v_type != VREG) 891 return (EIO); 892 mtx_lock(&np->n_mtx); 893 if (np->n_flag & NWRITEERR) { 894 np->n_flag &= ~NWRITEERR; 895 mtx_unlock(&np->n_mtx); 896 return (np->n_error); 897 } else 898 mtx_unlock(&np->n_mtx); 899 mtx_lock(&nmp->nm_mtx); 900 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 901 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 902 mtx_unlock(&nmp->nm_mtx); 903 (void)nfs_fsinfo(nmp, vp, cred, td); 904 } else 905 mtx_unlock(&nmp->nm_mtx); 906 907 /* 908 * Synchronously flush pending buffers if we are in synchronous 909 * mode or if we are appending. 910 */ 911 if (ioflag & (IO_APPEND | IO_SYNC)) { 912 mtx_lock(&np->n_mtx); 913 if (np->n_flag & NMODIFIED) { 914 mtx_unlock(&np->n_mtx); 915 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 916 /* 917 * Require non-blocking, synchronous writes to 918 * dirty files to inform the program it needs 919 * to fsync(2) explicitly. 920 */ 921 if (ioflag & IO_NDELAY) 922 return (EAGAIN); 923 #endif 924 flush_and_restart: 925 np->n_attrstamp = 0; 926 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 927 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 928 if (error) 929 return (error); 930 } else 931 mtx_unlock(&np->n_mtx); 932 } 933 934 /* 935 * If IO_APPEND then load uio_offset. We restart here if we cannot 936 * get the append lock. 937 */ 938 if (ioflag & IO_APPEND) { 939 np->n_attrstamp = 0; 940 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 941 error = VOP_GETATTR(vp, &vattr, cred); 942 if (error) 943 return (error); 944 mtx_lock(&np->n_mtx); 945 uio->uio_offset = np->n_size; 946 mtx_unlock(&np->n_mtx); 947 } 948 949 if (uio->uio_offset < 0) 950 return (EINVAL); 951 end = uio->uio_offset + uio->uio_resid; 952 if (end > nmp->nm_maxfilesize || end < uio->uio_offset) 953 return (EFBIG); 954 if (uio->uio_resid == 0) 955 return (0); 956 957 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 958 return nfs_directio_write(vp, uio, cred, ioflag); 959 960 /* 961 * Maybe this should be above the vnode op call, but so long as 962 * file servers have no limits, i don't think it matters 963 */ 964 if (vn_rlimit_fsize(vp, uio, td)) 965 return (EFBIG); 966 967 biosize = vp->v_bufobj.bo_bsize; 968 /* 969 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 970 * would exceed the local maximum per-file write commit size when 971 * combined with those, we must decide whether to flush, 972 * go synchronous, or return error. We don't bother checking 973 * IO_UNIT -- we just make all writes atomic anyway, as there's 974 * no point optimizing for something that really won't ever happen. 975 */ 976 if (!(ioflag & IO_SYNC)) { 977 int nflag; 978 979 mtx_lock(&np->n_mtx); 980 nflag = np->n_flag; 981 mtx_unlock(&np->n_mtx); 982 int needrestart = 0; 983 if (nmp->nm_wcommitsize < uio->uio_resid) { 984 /* 985 * If this request could not possibly be completed 986 * without exceeding the maximum outstanding write 987 * commit size, see if we can convert it into a 988 * synchronous write operation. 989 */ 990 if (ioflag & IO_NDELAY) 991 return (EAGAIN); 992 ioflag |= IO_SYNC; 993 if (nflag & NMODIFIED) 994 needrestart = 1; 995 } else if (nflag & NMODIFIED) { 996 int wouldcommit = 0; 997 BO_LOCK(&vp->v_bufobj); 998 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 999 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1000 b_bobufs) { 1001 if (bp->b_flags & B_NEEDCOMMIT) 1002 wouldcommit += bp->b_bcount; 1003 } 1004 } 1005 BO_UNLOCK(&vp->v_bufobj); 1006 /* 1007 * Since we're not operating synchronously and 1008 * bypassing the buffer cache, we are in a commit 1009 * and holding all of these buffers whether 1010 * transmitted or not. If not limited, this 1011 * will lead to the buffer cache deadlocking, 1012 * as no one else can flush our uncommitted buffers. 1013 */ 1014 wouldcommit += uio->uio_resid; 1015 /* 1016 * If we would initially exceed the maximum 1017 * outstanding write commit size, flush and restart. 1018 */ 1019 if (wouldcommit > nmp->nm_wcommitsize) 1020 needrestart = 1; 1021 } 1022 if (needrestart) 1023 goto flush_and_restart; 1024 } 1025 1026 do { 1027 nfsstats.biocache_writes++; 1028 lbn = uio->uio_offset / biosize; 1029 on = uio->uio_offset & (biosize-1); 1030 n = min((unsigned)(biosize - on), uio->uio_resid); 1031 again: 1032 /* 1033 * Handle direct append and file extension cases, calculate 1034 * unaligned buffer size. 1035 */ 1036 mtx_lock(&np->n_mtx); 1037 if (uio->uio_offset == np->n_size && n) { 1038 mtx_unlock(&np->n_mtx); 1039 /* 1040 * Get the buffer (in its pre-append state to maintain 1041 * B_CACHE if it was previously set). Resize the 1042 * nfsnode after we have locked the buffer to prevent 1043 * readers from reading garbage. 1044 */ 1045 bcount = on; 1046 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1047 1048 if (bp != NULL) { 1049 long save; 1050 1051 mtx_lock(&np->n_mtx); 1052 np->n_size = uio->uio_offset + n; 1053 np->n_flag |= NMODIFIED; 1054 vnode_pager_setsize(vp, np->n_size); 1055 mtx_unlock(&np->n_mtx); 1056 1057 save = bp->b_flags & B_CACHE; 1058 bcount += n; 1059 allocbuf(bp, bcount); 1060 bp->b_flags |= save; 1061 } 1062 } else { 1063 /* 1064 * Obtain the locked cache block first, and then 1065 * adjust the file's size as appropriate. 1066 */ 1067 bcount = on + n; 1068 if ((off_t)lbn * biosize + bcount < np->n_size) { 1069 if ((off_t)(lbn + 1) * biosize < np->n_size) 1070 bcount = biosize; 1071 else 1072 bcount = np->n_size - (off_t)lbn * biosize; 1073 } 1074 mtx_unlock(&np->n_mtx); 1075 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1076 mtx_lock(&np->n_mtx); 1077 if (uio->uio_offset + n > np->n_size) { 1078 np->n_size = uio->uio_offset + n; 1079 np->n_flag |= NMODIFIED; 1080 vnode_pager_setsize(vp, np->n_size); 1081 } 1082 mtx_unlock(&np->n_mtx); 1083 } 1084 1085 if (!bp) { 1086 error = nfs_sigintr(nmp, td); 1087 if (!error) 1088 error = EINTR; 1089 break; 1090 } 1091 1092 /* 1093 * Issue a READ if B_CACHE is not set. In special-append 1094 * mode, B_CACHE is based on the buffer prior to the write 1095 * op and is typically set, avoiding the read. If a read 1096 * is required in special append mode, the server will 1097 * probably send us a short-read since we extended the file 1098 * on our end, resulting in b_resid == 0 and, thusly, 1099 * B_CACHE getting set. 1100 * 1101 * We can also avoid issuing the read if the write covers 1102 * the entire buffer. We have to make sure the buffer state 1103 * is reasonable in this case since we will not be initiating 1104 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1105 * more information. 1106 * 1107 * B_CACHE may also be set due to the buffer being cached 1108 * normally. 1109 */ 1110 1111 if (on == 0 && n == bcount) { 1112 bp->b_flags |= B_CACHE; 1113 bp->b_flags &= ~B_INVAL; 1114 bp->b_ioflags &= ~BIO_ERROR; 1115 } 1116 1117 if ((bp->b_flags & B_CACHE) == 0) { 1118 bp->b_iocmd = BIO_READ; 1119 vfs_busy_pages(bp, 0); 1120 error = nfs_doio(vp, bp, cred, td); 1121 if (error) { 1122 brelse(bp); 1123 break; 1124 } 1125 } 1126 if (bp->b_wcred == NOCRED) 1127 bp->b_wcred = crhold(cred); 1128 mtx_lock(&np->n_mtx); 1129 np->n_flag |= NMODIFIED; 1130 mtx_unlock(&np->n_mtx); 1131 1132 /* 1133 * If dirtyend exceeds file size, chop it down. This should 1134 * not normally occur but there is an append race where it 1135 * might occur XXX, so we log it. 1136 * 1137 * If the chopping creates a reverse-indexed or degenerate 1138 * situation with dirtyoff/end, we 0 both of them. 1139 */ 1140 1141 if (bp->b_dirtyend > bcount) { 1142 nfs_printf("NFS append race @%lx:%d\n", 1143 (long)bp->b_blkno * DEV_BSIZE, 1144 bp->b_dirtyend - bcount); 1145 bp->b_dirtyend = bcount; 1146 } 1147 1148 if (bp->b_dirtyoff >= bp->b_dirtyend) 1149 bp->b_dirtyoff = bp->b_dirtyend = 0; 1150 1151 /* 1152 * If the new write will leave a contiguous dirty 1153 * area, just update the b_dirtyoff and b_dirtyend, 1154 * otherwise force a write rpc of the old dirty area. 1155 * 1156 * While it is possible to merge discontiguous writes due to 1157 * our having a B_CACHE buffer ( and thus valid read data 1158 * for the hole), we don't because it could lead to 1159 * significant cache coherency problems with multiple clients, 1160 * especially if locking is implemented later on. 1161 * 1162 * as an optimization we could theoretically maintain 1163 * a linked list of discontinuous areas, but we would still 1164 * have to commit them separately so there isn't much 1165 * advantage to it except perhaps a bit of asynchronization. 1166 */ 1167 1168 if (bp->b_dirtyend > 0 && 1169 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1170 if (bwrite(bp) == EINTR) { 1171 error = EINTR; 1172 break; 1173 } 1174 goto again; 1175 } 1176 1177 error = uiomove((char *)bp->b_data + on, n, uio); 1178 1179 /* 1180 * Since this block is being modified, it must be written 1181 * again and not just committed. Since write clustering does 1182 * not work for the stage 1 data write, only the stage 2 1183 * commit rpc, we have to clear B_CLUSTEROK as well. 1184 */ 1185 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1186 1187 if (error) { 1188 bp->b_ioflags |= BIO_ERROR; 1189 brelse(bp); 1190 break; 1191 } 1192 1193 /* 1194 * Only update dirtyoff/dirtyend if not a degenerate 1195 * condition. 1196 */ 1197 if (n) { 1198 if (bp->b_dirtyend > 0) { 1199 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1200 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1201 } else { 1202 bp->b_dirtyoff = on; 1203 bp->b_dirtyend = on + n; 1204 } 1205 vfs_bio_set_valid(bp, on, n); 1206 } 1207 1208 /* 1209 * If IO_SYNC do bwrite(). 1210 * 1211 * IO_INVAL appears to be unused. The idea appears to be 1212 * to turn off caching in this case. Very odd. XXX 1213 */ 1214 if ((ioflag & IO_SYNC)) { 1215 if (ioflag & IO_INVAL) 1216 bp->b_flags |= B_NOCACHE; 1217 error = bwrite(bp); 1218 if (error) 1219 break; 1220 } else if ((n + on) == biosize) { 1221 bp->b_flags |= B_ASYNC; 1222 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1223 } else { 1224 bdwrite(bp); 1225 } 1226 } while (uio->uio_resid > 0 && n > 0); 1227 1228 return (error); 1229 } 1230 1231 /* 1232 * Get an nfs cache block. 1233 * 1234 * Allocate a new one if the block isn't currently in the cache 1235 * and return the block marked busy. If the calling process is 1236 * interrupted by a signal for an interruptible mount point, return 1237 * NULL. 1238 * 1239 * The caller must carefully deal with the possible B_INVAL state of 1240 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1241 * indirectly), so synchronous reads can be issued without worrying about 1242 * the B_INVAL state. We have to be a little more careful when dealing 1243 * with writes (see comments in nfs_write()) when extending a file past 1244 * its EOF. 1245 */ 1246 static struct buf * 1247 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1248 { 1249 struct buf *bp; 1250 struct mount *mp; 1251 struct nfsmount *nmp; 1252 1253 mp = vp->v_mount; 1254 nmp = VFSTONFS(mp); 1255 1256 if (nmp->nm_flag & NFSMNT_INT) { 1257 sigset_t oldset; 1258 1259 nfs_set_sigmask(td, &oldset); 1260 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1261 nfs_restore_sigmask(td, &oldset); 1262 while (bp == NULL) { 1263 if (nfs_sigintr(nmp, td)) 1264 return (NULL); 1265 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1266 } 1267 } else { 1268 bp = getblk(vp, bn, size, 0, 0, 0); 1269 } 1270 1271 if (vp->v_type == VREG) 1272 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1273 return (bp); 1274 } 1275 1276 /* 1277 * Flush and invalidate all dirty buffers. If another process is already 1278 * doing the flush, just wait for completion. 1279 */ 1280 int 1281 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1282 { 1283 struct nfsnode *np = VTONFS(vp); 1284 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1285 int error = 0, slpflag, slptimeo; 1286 int old_lock = 0; 1287 1288 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1289 1290 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1291 intrflg = 0; 1292 if (intrflg) { 1293 slpflag = NFS_PCATCH; 1294 slptimeo = 2 * hz; 1295 } else { 1296 slpflag = 0; 1297 slptimeo = 0; 1298 } 1299 1300 old_lock = nfs_upgrade_vnlock(vp); 1301 if (vp->v_iflag & VI_DOOMED) { 1302 /* 1303 * Since vgonel() uses the generic vinvalbuf() to flush 1304 * dirty buffers and it does not call this function, it 1305 * is safe to just return OK when VI_DOOMED is set. 1306 */ 1307 nfs_downgrade_vnlock(vp, old_lock); 1308 return (0); 1309 } 1310 1311 /* 1312 * Now, flush as required. 1313 */ 1314 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1315 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1316 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1317 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1318 /* 1319 * If the page clean was interrupted, fail the invalidation. 1320 * Not doing so, we run the risk of losing dirty pages in the 1321 * vinvalbuf() call below. 1322 */ 1323 if (intrflg && (error = nfs_sigintr(nmp, td))) 1324 goto out; 1325 } 1326 1327 error = vinvalbuf(vp, flags, slpflag, 0); 1328 while (error) { 1329 if (intrflg && (error = nfs_sigintr(nmp, td))) 1330 goto out; 1331 error = vinvalbuf(vp, flags, 0, slptimeo); 1332 } 1333 mtx_lock(&np->n_mtx); 1334 if (np->n_directio_asyncwr == 0) 1335 np->n_flag &= ~NMODIFIED; 1336 mtx_unlock(&np->n_mtx); 1337 out: 1338 nfs_downgrade_vnlock(vp, old_lock); 1339 return error; 1340 } 1341 1342 /* 1343 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1344 * This is mainly to avoid queueing async I/O requests when the nfsiods 1345 * are all hung on a dead server. 1346 * 1347 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1348 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1349 */ 1350 int 1351 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1352 { 1353 int iod; 1354 int gotiod; 1355 int slpflag = 0; 1356 int slptimeo = 0; 1357 int error, error2; 1358 1359 /* 1360 * Commits are usually short and sweet so lets save some cpu and 1361 * leave the async daemons for more important rpc's (such as reads 1362 * and writes). 1363 * 1364 * Readdirplus RPCs do vget()s to acquire the vnodes for entries 1365 * in the directory in order to update attributes. This can deadlock 1366 * with another thread that is waiting for async I/O to be done by 1367 * an nfsiod thread while holding a lock on one of these vnodes. 1368 * To avoid this deadlock, don't allow the async nfsiod threads to 1369 * perform Readdirplus RPCs. 1370 */ 1371 mtx_lock(&nfs_iod_mtx); 1372 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1373 (nmp->nm_bufqiods > nfs_numasync / 2)) || 1374 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { 1375 mtx_unlock(&nfs_iod_mtx); 1376 return(EIO); 1377 } 1378 again: 1379 if (nmp->nm_flag & NFSMNT_INT) 1380 slpflag = NFS_PCATCH; 1381 gotiod = FALSE; 1382 1383 /* 1384 * Find a free iod to process this request. 1385 */ 1386 for (iod = 0; iod < nfs_numasync; iod++) 1387 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) { 1388 gotiod = TRUE; 1389 break; 1390 } 1391 1392 /* 1393 * Try to create one if none are free. 1394 */ 1395 if (!gotiod) 1396 nfs_nfsiodnew(); 1397 else { 1398 /* 1399 * Found one, so wake it up and tell it which 1400 * mount to process. 1401 */ 1402 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1403 iod, nmp)); 1404 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1405 nfs_iodmount[iod] = nmp; 1406 nmp->nm_bufqiods++; 1407 wakeup(&nfs_iodwant[iod]); 1408 } 1409 1410 /* 1411 * If none are free, we may already have an iod working on this mount 1412 * point. If so, it will process our request. 1413 */ 1414 if (!gotiod) { 1415 if (nmp->nm_bufqiods > 0) { 1416 NFS_DPF(ASYNCIO, 1417 ("nfs_asyncio: %d iods are already processing mount %p\n", 1418 nmp->nm_bufqiods, nmp)); 1419 gotiod = TRUE; 1420 } 1421 } 1422 1423 /* 1424 * If we have an iod which can process the request, then queue 1425 * the buffer. 1426 */ 1427 if (gotiod) { 1428 /* 1429 * Ensure that the queue never grows too large. We still want 1430 * to asynchronize so we block rather then return EIO. 1431 */ 1432 while (nmp->nm_bufqlen >= 2 * nfs_numasync) { 1433 NFS_DPF(ASYNCIO, 1434 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1435 nmp->nm_bufqwant = TRUE; 1436 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1437 slpflag | PRIBIO, 1438 "nfsaio", slptimeo); 1439 if (error) { 1440 error2 = nfs_sigintr(nmp, td); 1441 if (error2) { 1442 mtx_unlock(&nfs_iod_mtx); 1443 return (error2); 1444 } 1445 if (slpflag == NFS_PCATCH) { 1446 slpflag = 0; 1447 slptimeo = 2 * hz; 1448 } 1449 } 1450 /* 1451 * We might have lost our iod while sleeping, 1452 * so check and loop if nescessary. 1453 */ 1454 goto again; 1455 } 1456 1457 /* We might have lost our nfsiod */ 1458 if (nmp->nm_bufqiods == 0) { 1459 NFS_DPF(ASYNCIO, 1460 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1461 goto again; 1462 } 1463 1464 if (bp->b_iocmd == BIO_READ) { 1465 if (bp->b_rcred == NOCRED && cred != NOCRED) 1466 bp->b_rcred = crhold(cred); 1467 } else { 1468 if (bp->b_wcred == NOCRED && cred != NOCRED) 1469 bp->b_wcred = crhold(cred); 1470 } 1471 1472 if (bp->b_flags & B_REMFREE) 1473 bremfreef(bp); 1474 BUF_KERNPROC(bp); 1475 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1476 nmp->nm_bufqlen++; 1477 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1478 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1479 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1480 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1481 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1482 } 1483 mtx_unlock(&nfs_iod_mtx); 1484 return (0); 1485 } 1486 1487 mtx_unlock(&nfs_iod_mtx); 1488 1489 /* 1490 * All the iods are busy on other mounts, so return EIO to 1491 * force the caller to process the i/o synchronously. 1492 */ 1493 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1494 return (EIO); 1495 } 1496 1497 void 1498 nfs_doio_directwrite(struct buf *bp) 1499 { 1500 int iomode, must_commit; 1501 struct uio *uiop = (struct uio *)bp->b_caller1; 1502 char *iov_base = uiop->uio_iov->iov_base; 1503 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1504 1505 iomode = NFSV3WRITE_FILESYNC; 1506 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1507 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1508 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1509 free(iov_base, M_NFSDIRECTIO); 1510 free(uiop->uio_iov, M_NFSDIRECTIO); 1511 free(uiop, M_NFSDIRECTIO); 1512 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1513 struct nfsnode *np = VTONFS(bp->b_vp); 1514 mtx_lock(&np->n_mtx); 1515 np->n_directio_asyncwr--; 1516 if (np->n_directio_asyncwr == 0) { 1517 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1518 if ((np->n_flag & NFSYNCWAIT)) { 1519 np->n_flag &= ~NFSYNCWAIT; 1520 wakeup((caddr_t)&np->n_directio_asyncwr); 1521 } 1522 } 1523 mtx_unlock(&np->n_mtx); 1524 } 1525 bp->b_vp = NULL; 1526 relpbuf(bp, &nfs_pbuf_freecnt); 1527 } 1528 1529 /* 1530 * Do an I/O operation to/from a cache block. This may be called 1531 * synchronously or from an nfsiod. 1532 */ 1533 int 1534 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1535 { 1536 struct uio *uiop; 1537 struct nfsnode *np; 1538 struct nfsmount *nmp; 1539 int error = 0, iomode, must_commit = 0; 1540 struct uio uio; 1541 struct iovec io; 1542 struct proc *p = td ? td->td_proc : NULL; 1543 uint8_t iocmd; 1544 1545 np = VTONFS(vp); 1546 nmp = VFSTONFS(vp->v_mount); 1547 uiop = &uio; 1548 uiop->uio_iov = &io; 1549 uiop->uio_iovcnt = 1; 1550 uiop->uio_segflg = UIO_SYSSPACE; 1551 uiop->uio_td = td; 1552 1553 /* 1554 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1555 * do this here so we do not have to do it in all the code that 1556 * calls us. 1557 */ 1558 bp->b_flags &= ~B_INVAL; 1559 bp->b_ioflags &= ~BIO_ERROR; 1560 1561 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1562 iocmd = bp->b_iocmd; 1563 if (iocmd == BIO_READ) { 1564 io.iov_len = uiop->uio_resid = bp->b_bcount; 1565 io.iov_base = bp->b_data; 1566 uiop->uio_rw = UIO_READ; 1567 1568 switch (vp->v_type) { 1569 case VREG: 1570 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1571 nfsstats.read_bios++; 1572 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1573 1574 if (!error) { 1575 if (uiop->uio_resid) { 1576 /* 1577 * If we had a short read with no error, we must have 1578 * hit a file hole. We should zero-fill the remainder. 1579 * This can also occur if the server hits the file EOF. 1580 * 1581 * Holes used to be able to occur due to pending 1582 * writes, but that is not possible any longer. 1583 */ 1584 int nread = bp->b_bcount - uiop->uio_resid; 1585 int left = uiop->uio_resid; 1586 1587 if (left > 0) 1588 bzero((char *)bp->b_data + nread, left); 1589 uiop->uio_resid = 0; 1590 } 1591 } 1592 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1593 if (p && (vp->v_vflag & VV_TEXT)) { 1594 mtx_lock(&np->n_mtx); 1595 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1596 mtx_unlock(&np->n_mtx); 1597 PROC_LOCK(p); 1598 killproc(p, "text file modification"); 1599 PROC_UNLOCK(p); 1600 } else 1601 mtx_unlock(&np->n_mtx); 1602 } 1603 break; 1604 case VLNK: 1605 uiop->uio_offset = (off_t)0; 1606 nfsstats.readlink_bios++; 1607 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1608 break; 1609 case VDIR: 1610 nfsstats.readdir_bios++; 1611 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1612 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1613 error = nfs_readdirplusrpc(vp, uiop, cr); 1614 if (error == NFSERR_NOTSUPP) 1615 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1616 } 1617 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1618 error = nfs_readdirrpc(vp, uiop, cr); 1619 /* 1620 * end-of-directory sets B_INVAL but does not generate an 1621 * error. 1622 */ 1623 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1624 bp->b_flags |= B_INVAL; 1625 break; 1626 default: 1627 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1628 break; 1629 }; 1630 if (error) { 1631 bp->b_ioflags |= BIO_ERROR; 1632 bp->b_error = error; 1633 } 1634 } else { 1635 /* 1636 * If we only need to commit, try to commit 1637 */ 1638 if (bp->b_flags & B_NEEDCOMMIT) { 1639 int retv; 1640 off_t off; 1641 1642 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1643 retv = (nmp->nm_rpcops->nr_commit)( 1644 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1645 bp->b_wcred, td); 1646 if (retv == 0) { 1647 bp->b_dirtyoff = bp->b_dirtyend = 0; 1648 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1649 bp->b_resid = 0; 1650 bufdone(bp); 1651 return (0); 1652 } 1653 if (retv == NFSERR_STALEWRITEVERF) { 1654 nfs_clearcommit(vp->v_mount); 1655 } 1656 } 1657 1658 /* 1659 * Setup for actual write 1660 */ 1661 mtx_lock(&np->n_mtx); 1662 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1663 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1664 mtx_unlock(&np->n_mtx); 1665 1666 if (bp->b_dirtyend > bp->b_dirtyoff) { 1667 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1668 - bp->b_dirtyoff; 1669 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1670 + bp->b_dirtyoff; 1671 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1672 uiop->uio_rw = UIO_WRITE; 1673 nfsstats.write_bios++; 1674 1675 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1676 iomode = NFSV3WRITE_UNSTABLE; 1677 else 1678 iomode = NFSV3WRITE_FILESYNC; 1679 1680 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1681 1682 /* 1683 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1684 * to cluster the buffers needing commit. This will allow 1685 * the system to submit a single commit rpc for the whole 1686 * cluster. We can do this even if the buffer is not 100% 1687 * dirty (relative to the NFS blocksize), so we optimize the 1688 * append-to-file-case. 1689 * 1690 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1691 * cleared because write clustering only works for commit 1692 * rpc's, not for the data portion of the write). 1693 */ 1694 1695 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1696 bp->b_flags |= B_NEEDCOMMIT; 1697 if (bp->b_dirtyoff == 0 1698 && bp->b_dirtyend == bp->b_bcount) 1699 bp->b_flags |= B_CLUSTEROK; 1700 } else { 1701 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1702 } 1703 1704 /* 1705 * For an interrupted write, the buffer is still valid 1706 * and the write hasn't been pushed to the server yet, 1707 * so we can't set BIO_ERROR and report the interruption 1708 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1709 * is not relevant, so the rpc attempt is essentially 1710 * a noop. For the case of a V3 write rpc not being 1711 * committed to stable storage, the block is still 1712 * dirty and requires either a commit rpc or another 1713 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1714 * the block is reused. This is indicated by setting 1715 * the B_DELWRI and B_NEEDCOMMIT flags. 1716 * 1717 * If the buffer is marked B_PAGING, it does not reside on 1718 * the vp's paging queues so we cannot call bdirty(). The 1719 * bp in this case is not an NFS cache block so we should 1720 * be safe. XXX 1721 * 1722 * The logic below breaks up errors into recoverable and 1723 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1724 * and keep the buffer around for potential write retries. 1725 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1726 * and save the error in the nfsnode. This is less than ideal 1727 * but necessary. Keeping such buffers around could potentially 1728 * cause buffer exhaustion eventually (they can never be written 1729 * out, so will get constantly be re-dirtied). It also causes 1730 * all sorts of vfs panics. For non-recoverable write errors, 1731 * also invalidate the attrcache, so we'll be forced to go over 1732 * the wire for this object, returning an error to user on next 1733 * call (most of the time). 1734 */ 1735 if (error == EINTR || error == EIO || error == ETIMEDOUT 1736 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1737 int s; 1738 1739 s = splbio(); 1740 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1741 if ((bp->b_flags & B_PAGING) == 0) { 1742 bdirty(bp); 1743 bp->b_flags &= ~B_DONE; 1744 } 1745 if (error && (bp->b_flags & B_ASYNC) == 0) 1746 bp->b_flags |= B_EINTR; 1747 splx(s); 1748 } else { 1749 if (error) { 1750 bp->b_ioflags |= BIO_ERROR; 1751 bp->b_flags |= B_INVAL; 1752 bp->b_error = np->n_error = error; 1753 mtx_lock(&np->n_mtx); 1754 np->n_flag |= NWRITEERR; 1755 np->n_attrstamp = 0; 1756 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1757 mtx_unlock(&np->n_mtx); 1758 } 1759 bp->b_dirtyoff = bp->b_dirtyend = 0; 1760 } 1761 } else { 1762 bp->b_resid = 0; 1763 bufdone(bp); 1764 return (0); 1765 } 1766 } 1767 bp->b_resid = uiop->uio_resid; 1768 if (must_commit) 1769 nfs_clearcommit(vp->v_mount); 1770 bufdone(bp); 1771 return (error); 1772 } 1773 1774 /* 1775 * Used to aid in handling ftruncate() operations on the NFS client side. 1776 * Truncation creates a number of special problems for NFS. We have to 1777 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1778 * we have to properly handle VM pages or (potentially dirty) buffers 1779 * that straddle the truncation point. 1780 */ 1781 1782 int 1783 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1784 { 1785 struct nfsnode *np = VTONFS(vp); 1786 u_quad_t tsize; 1787 int biosize = vp->v_bufobj.bo_bsize; 1788 int error = 0; 1789 1790 mtx_lock(&np->n_mtx); 1791 tsize = np->n_size; 1792 np->n_size = nsize; 1793 mtx_unlock(&np->n_mtx); 1794 1795 if (nsize < tsize) { 1796 struct buf *bp; 1797 daddr_t lbn; 1798 int bufsize; 1799 1800 /* 1801 * vtruncbuf() doesn't get the buffer overlapping the 1802 * truncation point. We may have a B_DELWRI and/or B_CACHE 1803 * buffer that now needs to be truncated. 1804 */ 1805 error = vtruncbuf(vp, cred, td, nsize, biosize); 1806 lbn = nsize / biosize; 1807 bufsize = nsize & (biosize - 1); 1808 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1809 if (!bp) 1810 return EINTR; 1811 if (bp->b_dirtyoff > bp->b_bcount) 1812 bp->b_dirtyoff = bp->b_bcount; 1813 if (bp->b_dirtyend > bp->b_bcount) 1814 bp->b_dirtyend = bp->b_bcount; 1815 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1816 brelse(bp); 1817 } else { 1818 vnode_pager_setsize(vp, nsize); 1819 } 1820 return(error); 1821 } 1822
Cache object: 9aeaa3e69dfaa4be87218d5541d1a1de
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]